NEWMindMap Digital has acquired Bluetide.co— deepening our data & agentic-AI stack.Read more →
Home · Services · Chatbots & NLP
Conversational AI · NLP Engineering

Conversations that resolve, not just deflect

Most chatbots succeed at one metric and one metric only: they make the contact-volume report look better while customer satisfaction falls. We build conversational systems with a different definition of success — actual resolution of the customer's issue, end-to-end, including the booking, the payment, the status update, the escalation. The bot does the work, not the deflection.

67%
Avg first-contact resolution
12 langs
Including African and Gulf dialects
94%
Intent classification accuracy
48h
Standard channel deployment
67%
First-contact resolution
74%
Containment rate
4.4/5
CSAT on bot conversations
12 sec
Time to action completion
Capabilities

What we deliver

Channels as a feature, not a project

One bot serving WhatsApp Business, the web widget, Telegram, Facebook Messenger, Microsoft Teams, voice IVR, and email — with a single intent model, a single knowledge base, and channel-specific rendering handled by the platform. Adding a new channel is a configuration change, not a re-implementation.

RAG-grounded knowledge

Answers grounded in your private knowledge base — product documents, policies, FAQs, regulatory guidance — with retrieval that returns source paragraphs as citations. No hallucinations on factual questions, with an answerability classifier that escalates rather than confabulates when the knowledge base does not cover the question.

Action-taking agents

The bot calls your APIs to actually do things: book the appointment, update the address, raise the dispute, restart the modem, transfer the funds within limits, raise the ticket with full context. Authentication is delegated to your identity provider, audit logs flow to your SIEM, and the bot has only the permissions you grant it.

Multilingual including the hard languages

Twelve languages out of the box including Arabic with Gulf and Levantine dialects, English, French, Swahili, Hausa, Yoruba, Hindi, Urdu, Bengali, Tamil, Bahasa, and Portuguese. We train on your historical conversation logs so the bot speaks like your customers, not like a US translation engine.

12+ production languages

Human handoff with full context

When the bot escalates, the agent inherits the full conversation transcript, the customer's identified intent, the actions already taken, the customer's identity from your IDP, and the suggested next-best action. Zendesk, Freshdesk, Salesforce Service Cloud, ServiceNow, or your custom CRM — pre-built connectors for the common ones, REST for the rest.

Conversation intelligence

Intent heatmaps, deflection and resolution rates, CSAT correlated to conversation patterns, prompt-injection and abuse detection, and unresolved-question mining that feeds your knowledge-base team. The platform tells you which content to write next based on what customers actually asked.

Live Demo

WhatsApp banking bot

ChatNext — WhatsApp Banking Bot
What's my account balance?
Reference Architecture

How a query actually flows.

A real trace through the sovereign stack. Six stages, ~1.4 seconds end-to-end, zero packets leaving your perimeter.

QUERY TRACE · LIVEtrace_id 0x8c41a2b9usr_4821
SOVEREIGN · ON-PREM·17:42:09 IST·● 200 OK
01
User submit
"Q3 underwriting flags"
42ms
02
Embed
bge-large-en · 1024d
180ms
03
Vector search
pgvector · k=32
90ms
04
Rerank · guardrail
PII · safety · top-8
140ms
05
Sovereign LLM
Llama 3.1 · 70B · local
940ms
06
Compose · cite
8 docs · markdown
28ms
WATERFALL · LAST QUERYtotal 1.42s · sla < 2s
USER SUBMIT
42 ms
EMBED · bge
180 ms
VECTOR SEARCH
90 ms
RERANK · GUARD
140 ms
LLM INFERENCE
940 ms
COMPOSE · CITE
28 ms
0 ms50010001500
RESPONSE · SAMPLE8 docs cited · 99% confidence
Q"Summarise Q3 underwriting flags"
A3 anomalies detected in Q3 underwriting [1]: velocity spikes in segment-NA [4], policy concentration above threshold [7], and 2 dormant accounts re-activated [11].
[1]q3_uw_summary.pdf
[4]region_na_h2.xlsx
[7]concentration_log.csv
[11]dormant_audit.pdf
LIVE TRACES · LAST 90s12 ok · 0 failed · 0 egress
17:42:090x8c41a2b9usr_4821rag.query8 docs · llama-70b1.42 s● OK
17:42:040x8c419f44svc_kycllm.classifydoc=invoice · 99%0.81 s● OK
17:41:580x8c419b10usr_2110agent.runfraud_check · 12 rules2.04 s● OK
17:41:510x8c41960cusr_4821rag.query6 docs · llama-70b1.11 s● OK
17:41:460x8c4192e8svc_ocrllm.extract12 fields · 98.6%0.94 s● OK
17:41:390x8c418f10usr_8801agent.rununderwrite · pass1.66 s● OK
ZERO API EGRESS · 0 BYTES OUTALL STAGES INSIDE PERIMETEREVERY TRACE WRITTEN TO YOUR AUDIT STORE↗ SOVEREIGN
Methodology

How we deliver

01

Top-intent mapping

We analyse three to six months of historical contact logs to identify the top twenty intents by volume, the top ten by handling cost, and the top five by customer pain. The intersection becomes the launch backlog. We will not let you start with a bot that answers questions nobody is asking.

02

Build and integrate

Two to four weeks to implement the launch intents, connect the knowledge base, integrate the action APIs, and configure the channel surfaces. Every action is end-to-end tested against your sandbox systems before any customer sees it.

03

Shadow on real traffic

We run the bot in shadow mode against live customer conversations for two weeks, scoring intent classification, answer accuracy, and would-be-resolution rate without affecting actual customers. Misclassifications and gaps get fixed before launch.

04

Controlled launch

Deploy to a single channel and a subset of customers, monitor every conversation, and ramp up over four to six weeks. War-room support from MindMap for go-live and the first month of hypercare.

05

Continuous expansion

Sprint cadence adds intents, languages, channels, and integrations. The platform improves automatically from production traffic — misclassifications you label become training data, unanswered questions feed the knowledge-base team, and the resolution rate trends up quarter on quarter.

Technology

The stack we build on

NLP engine
Rasa Open Source
LangChain
LangGraph
spaCy NER
Multilingual BERT
Custom LLM fine-tunes
Channels
WhatsApp Business API
Twilio Programmable Voice
Telegram Bot API
Facebook Messenger
Microsoft Teams
Web SDK
Knowledge and retrieval
RAG pipelines
Qdrant / Milvus
OpenAI / Cohere embeddings
Knowledge graphs
FAQ extractors
Auto-curation
Integrations and analytics
Salesforce Service Cloud
Zendesk
Freshdesk
ServiceNow
Conversation analytics
A/B testing
"We launched on WhatsApp on a Friday with eight intents. Six months later it handles sixty-seven percent of our inbound volume across four channels and twenty-three intents, in English and Swahili, and our NPS has gone up eighteen points. The bot does not deflect — it resolves."
Head of Digital, Tier-1 East African Bank
Engagement Options

How we work together

ChatNext SaaS

Deploy our production-tested NLP platform — already running across three continents for fifty-plus banks and telcos — within forty-eight hours of contract. Includes channels, knowledge base, integration framework, and operations. Best for teams that want speed and proven scale.

Custom build on open source

Bespoke conversational AI on Rasa, LangChain, or LangGraph for requirements that need full source-code ownership, unique integrations, or specialised compliance constraints. Longer build but maximum flexibility. Typically eight to twelve weeks to first launch.

Hybrid

ChatNext as the foundation for the common intents, custom modules built alongside it for the specialised journeys. The most common pattern for large enterprises — speed where speed matters, custom where custom matters.

FAQ

Common questions

What is the difference between ChatNext and a custom build?+

ChatNext is our production-tested platform with four years of deployment across banking, telecoms, and retail — fifty-plus enterprise installations, three million-plus monthly conversations, twelve languages. A custom build on Rasa or LangChain makes sense when your requirements are highly specific or you need full source-code ownership. Most clients start with ChatNext and add custom modules where they need them — a hybrid pattern that gives speed without sacrificing control.

How do you handle languages that are poorly supported by mainstream NLP?+

We have built and fine-tuned models for African languages — Swahili, Hausa, Yoruba, Amharic — and dialectal Gulf and Levantine Arabic on training data drawn from our clients' actual conversation logs. The base capability is multilingual BERT or a Llama derivative, the fine-tunes are bespoke per language. For a new language we typically need twenty to fifty thousand labelled examples to reach production accuracy, which we draw from your historical contacts.

Can the bot actually take actions in our core systems?+

Yes — this is the design point. Our bots are agentic: they call your APIs to update records, raise tickets, book appointments, transfer funds within policy limits, and send confirmations. Authentication is delegated to your identity provider so the bot only has the customer's permissions; audit logs flow to your SIEM; rate limits and circuit breakers protect your downstream systems. Deflection without resolution is just an obstacle course — we will not build that.

How do you measure success beyond containment?+

Containment is the vanity metric. We measure first-contact resolution — did the customer's actual problem get solved without escalation — alongside post-bot CSAT, action-completion time, and downstream contact rate. Bots that artificially contain volume by frustrating customers into hanging up show up as high containment with high downstream contact and falling CSAT; the platform surfaces this pattern and flags it for tuning.

What happens with prompt injection and abuse?+

Layered defences. Input classifiers detect jailbreak attempts and obvious abuse before the prompt reaches the LLM. Output validators check responses against policy before delivery. Rate limits and behavioural anomaly detection catch automated abuse. A red-team test suite — built collaboratively with your security team — runs against every release. We document the residual risk in your CISO's language, not in marketing speak.

How long does deployment really take?+

On ChatNext, forty-eight hours to a working bot on a single channel with three to five intents — useful for proof. Two to four weeks to a launch-ready bot with ten-to-fifteen intents, integrations, and tested escalation. Eight to twelve weeks to a scaled deployment across multiple channels and languages. Custom builds add four to eight weeks for the foundation. Anyone promising you a production-grade conversational AI in a week is selling a demo.

Ready to explore Chatbots & NLP?

Speak to our engineering team. No sales pitch — just a technical conversation.

Start a conversation →
Talk to the product team